Efficient Micro-Structured Weight Unification and Pruning for Neural Network Compression